Submodularity in Input Node Selection for Networked Systems Efficient Algorithms for Performance and Controllability Andrew Clark, Basel Alomair, Linda Bushnell, and Radha Poovendran Networked systems are systems of interconnected components, in which the dynamics of each component (often referred to as a node or agent) are influenced by the behavior of neighboring components. Examples of networked systems include biological networks, ranging in scale from the gene interactions within a single cell to food webs of an entire ecosystem, critical infrastructures such as power grids, transportation systems, the Internet, and social networks. The growing importance of such systems has led to an interest in control of networks to ensure performance, stability, robustness, and resilience [1], [2], [3], [4]. Indeed, over the past several decades, a variety of control-theoretic methods have been employed to better understand and control networked systems [5]. Moreover, new sub-disciplines of control theory have emerged including networked control systems (where plants, sensors, and actuators are connected by communication networks), and control of cyber-physical systems [6], [7], [8]. One approach to control a networked system is to exert control at a subset of nodes, often referred to as input nodes, leaders, driver nodes, or seed nodes, depending on the application domain [9], [10], [11]. The remaining nodes in the network can then be steered to reach a desired 1 arXiv:1605.09465v1 [cs.SY] 31 May 2016
64
Embed
Submodularity in Input Node Selection for Networked Systems · control networked systems [5]. Moreover, new sub-disciplines of control theory have emerged including networked control
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Submodularity in Input Node Selection for
Networked Systems
Efficient Algorithms for Performance and Controllability
Andrew Clark, Basel Alomair, Linda Bushnell, and Radha Poovendran
Networked systems are systems of interconnected components, in which the dynamics
of each component (often referred to as a node or agent) are influenced by the behavior of
neighboring components. Examples of networked systems include biological networks, ranging in
scale from the gene interactions within a single cell to food webs of an entire ecosystem, critical
infrastructures such as power grids, transportation systems, the Internet, and social networks.
The growing importance of such systems has led to an interest in control of networks to ensure
performance, stability, robustness, and resilience [1], [2], [3], [4]. Indeed, over the past several
decades, a variety of control-theoretic methods have been employed to better understand and
control networked systems [5]. Moreover, new sub-disciplines of control theory have emerged
including networked control systems (where plants, sensors, and actuators are connected by
communication networks), and control of cyber-physical systems [6], [7], [8].
One approach to control a networked system is to exert control at a subset of nodes, often
referred to as input nodes, leaders, driver nodes, or seed nodes, depending on the application
domain [9], [10], [11]. The remaining nodes in the network can then be steered to reach a desired
1
arX
iv:1
605.
0946
5v1
[cs
.SY
] 3
1 M
ay 2
016
state by exploiting the network interconnections. This approach provides a scalable alternative
to directly supplying an input signal to each network node, and is naturally applicable to diverse
problems such as steering a flock of unmanned vehicles from a formation leader [12], targeting a
set of genes for drug delivery [13], and selectively advertising towards high-influence individuals
in a marketing campaign [10], [14].
Under this control framework, the choice of where to exert control becomes a design
consideration alongside the choice of how to apply control at selected nodes. The choice of
input nodes has been shown to impact a variety of interrelated system properties, including
controllability and observability [15], [16], [17], [18]; robustness of the system to noise, failures,
and attacks [19], [20]; rate of convergence to a desired state [21], [22], [23]; and the amount
of energy required for control [24]. Enumerating all possible input sets in order to select an
optimal set would, however, be computationally infeasible for large-scale networks. The strict
performance, robustness, and controllability requirements of networked systems, together with
the computational challenges of combinatorial optimization, have motivated the investigation
of mathematical structures that can improve the speed and performance of input selection
algorithms.
This article presents submodular optimization approaches for input node selection in
networked systems. Submodularity is a property of set functions, analogous to concavity of
continuous functions, that enables the development of computationally tractable (polynomial-
time) algorithms with provable optimality bounds. The submodular structures discussed in this
article can be exploited to develop efficient input selection algorithms [25]. This article will
describe these structures and the resulting algorithms, as well as discuss open problems, and
2
show the practicality of submodular methods for control of networked systems.
The submodular approach to input selection is divided into two components, submod-
ularity for performance and submodularity for controllability. Submodularity for performance
refers to the ability of the system to reach the desired operating point in a timely fashion
and in the presence of noise, link and node outages, and adversarial attacks. For such metrics,
the submodular structure arise from connections between the network dynamics and diffusion
processes on the underlying network.
Controllability refers to the goal of ensuring that the networked system can be driven
to a desired operating point by controlling the input nodes. This article will show that
several aspects of the controllability problem exhibit submodular structure, including the rank
of the controllability Gramian, structural controllability (controllability analysis based on the
interconnection structure and physical invariants of the networked system), and the control effort
expended by optimal control. Each of these problem types, however, will require a fundamentally
different algorithm, and may exhibit differing optimality guarantees.
This article is organized as follows. Background on submodularity and matroids is given
first. The class of networked systems considered in this article are presented. A submodular
approach to optimizing performance, including ensuring robustness to noise and smooth con-
vergence, is presented next. Submodular optimization methods for controllability are discussed,
followed by techniques for joint performance and controllability. The article concludes with a
discussion of open problems and a summary of results.
3
Background on Submodularity
Submodularity is a property of set functions f : 2V → R, which take as input a subset
of a finite set V and output a real number. A function f is submodular if, for any sets S and T
with S ⊆ T ⊆ V , and any element v ∈ V \ T ,
f(S ∪ {v})− f(S) ≥ f(T ∪ {v})− f(T ). (1)
A function f is supermodular if −f is submodular. Eq. (1) can be interpreted as a diminishing
returns property, in which adding the element v to a set S has a larger incremental impact
than adding v to a superset T . This property is analogous to concavity of continuous functions.
Functions that are submodular can be efficiently optimized with provable optimality guarantees
under a variety of constraints.
One example of a problem with submodular structure is set cover, defined as follows.
Let A = {a1, . . . , am} denote a finite set, and let R1, . . . , Rn denote subsets of A. Define
V = {1, . . . , n}, and for any subset X ⊆ V , let f(X) = |∪i∈XRi|, where | · | denotes the
cardinality of a set. The increment f(S ∪ {v}) − f(S) from adding an element v is equal to
the number of elements that are contained in Rv but not in ∪i∈SRi. As the size of S grows, the
number of elements in Rv\∪i∈SRi decreases (Figure 1). Hence, the function f(S) is submodular
as a function of S.
An additional example of a problem that exhibits submodularity is shown in Figure 2.
Each grey cylinder in the figure represents a sensor node, which can monitor a fixed coverage
area. A subset of sensors is active at any time, with the number of active sensors limited by
battery constraints of each sensor. The goal of covering the entire rectangular region is captured
4
A1
A3
A4
n2
n1, n5 n4
n7
n3, n6,n8
A2
Figure 1. Submodularity of the set cover problem. Define A to be a finite set and R1, . . . , Rn to
be subsets of A. The function f(S) is given by f(X) =∣∣⋃
i∈X Ri
∣∣, where | · | denotes cardinality
of a set. In this example, A = {a1, . . . , a8}, R1, . . . , R4 are defined as shown, S = {1}, T =
{1, 3, 4}, and v = 2. All elements contained in v are already contained in T , and hence adding
v to T results in no incremental increase in f .
by the function f(S), which is equal to the total area covered by the sensors in set S. The
incremental increase in coverage area from adding n5 to the set of active sensors is larger for a
smaller set of active sensors, denoted S (Figure 2(a)) than for the larger set T shown in Figure
2(b).
A function f : 2V → R is monotone nondecreasing (respectively, nonincreasing) if,
for any S ⊆ T , f(S) ≤ f(T ) (respectively, f(S) ≥ f(T )). Not all submodular functions are
monotone, and vice versa; however, functions that are both monotone and submodular can be
optimized with improved optimality bounds [26], [27].
5
(a)
f(S): blueshadedarea
n1
n2
n3
n4n5
n6 n7
n1
n2
n3
n4n5
n6 n7
S={n1,n2,n3}
Areacoveredbyn5:Dashedline
(b)
f(T):blueshadedarea
T ={n1,n2,n3,n7}
Areacoveredbyn5:Dashedline
Figure 2. Example of submodularity. Each sensor (grey cylinder) has a coverage area, equal to
the area that the sensor can monitor. The goal is to monitor the rectangular region by selecting a
subset of active sensors. The function f(S) is equal to the total area covered by the set of sensors,
denoted S. Solid blue circles represent the coverage areas of sensors that are active. The dashed
lines indicate the coverage area of the sensor n5. (a) Incremental increase in coverage area from
adding n5 to the set of active sensors S = {n1, n2, n3}. (b) Incremental increase in coverage area
from adding n5 to the set T = {n1, n2, n3, n7}. The incremental increase f(T ∪ {n5}) − f(T )
from adding n5 to T is smaller than the incremental increase from adding n5 to S, since most
of n5’s coverage area is already covered by the sensors in T . Hence f is a submodular function.
Submodular functions have composition rules, analogous to composition of convex
functions, that are useful when proving submodularity [28]. A nonnegative weighted sum of
submodular functions f1(S), . . . , fm(S) is submodular as a function of S. For any monotone
submodular function f(S), the function g(S) = max {f(S), c} is submodular for any constant
6
c. Submodular functions satisfy a complementarity property: if f : 2V → R is a submodular
function, then the function f(S) defined by f(S) = f(V \ S) is submodular as well. Finally, if
f : 2V → R is a nonincreasing supermodular function and g : R → R is an increasing convex
function, then the composition h = g ◦ f is a supermodular function.
A variety of problems arising in machine learning, social networking, and game theory
have inherent submodular structure, which has led to increased research interest in submodular
optimization techniques. For details, see “Applications of Submodularity”.
Matroids
Matroids can be understood as generalizations of linear systems to discrete set sys-
tems. Matroids are sufficiently general to provide insights into graphs, matchings, and linear
systems [29], [30]. Matroid structure enables efficient solution or approximation of a variety
of combinatorial optimization problems. In particular, submodular maximization with matroid
constraints is known to have polynomial-time approximation algorithms [27].
A matroid M is defined by an ordered pair (V, I), where V is a finite set and I is a
collection of subsets of V . The collection of sets I must satisfy three properties:
(M1) ∅ ∈ I
(M2) Y ∈ I and X ⊆ Y implies X ∈ I
(M3) If X, Y ∈ I and |X| < |Y |, then there exists y ∈ Y \X such that (X ∪ {y}) ∈ I.
If X ∈ I, then X is said to be independent in M. The three properties of I can be interpreted
by an analogy to linear independence of a collection of vectors (Table I). Let V denote a set of
7
vectors. The empty set of vectors is trivially independent, satisfying the first property. If a set of
vectors is linearly independent, then any subset is also independent, thus satisfying the second
property. Finally, if two sets of vectors X and Y are linearly independent and |X| < |Y |, then
there must be at least one element in Y \ X that is not in the span of X , and hence can be
added to X while preserving independence.
Property Linear Space Matroid
Ground set Set of vectors V ⊂ Rn Finite set V
Independence Linear independence of a set X ⊂ V Set S ∈ I
Basis Linearly independent vectors in V Maximal independent set
that span V
Rank Rank(X) = rank of matrix with Rank function
column set X ρ(X) , max {|X ′| : X ′ ⊆ X,X ′ ∈ I}
TABLE I
PROPERTIES OF MATROIDS AND THE ANALOGOUS PROPERTIES OF LINEAR SYSTEMS.
A basis is a maximal independent set of the matroid. By property (M3), all bases have
the same cardinality; otherwise, if two bases X and Y satisfied |X| < |Y |, then an element from
Y could be added to X while preserving independence, contradicting maximality of X . In the
linear independence analogy, the bases correspond exactly to bases of the set of vectors V .
The rank function of a matroid M is defined by ρ : 2V → Z≥0, with
ρ(X) = max {|X ′| : X ′ ⊆ X and X ′ ∈ I}.
In words, the rank of X is the maximum-cardinality independent subset of X . In the linear
8
independence analogy, the rank of a set of vectors is equivalent to the dimension of the span of
those vectors, or the rank of the matrix with those vectors as the columns.
A class of matroids that will be useful in the controllability analysis is the transversal
matroids. Let U = {u1, . . . , un} denote a finite set, and let W1, . . . ,Wm denote a collection of
subsets of U . A transversal matroid M = (V, I) can be defined by setting V = {1, . . . ,m} and
X ∈ I ⇔ There exists a one-to-one mapping f : U → X with ui ∈ Wf(i).
The transversal matroid can best be interpreted using graph matchings (for details, see “Graph
Matchings”). A bipartite graph can be constructed with vertices indexed {u1, . . . , un} on the left,
vertices {w1, . . . , wm} on the right, and an edge (ui, wj) if i ∈ Wj . A set X ⊆ V is independent
if there is a matching in this bipartite graph in which {wi : i ∈ X} is matched (Figure 3).
Submodular Optimization Algorithms
Submodular structure enables the development of efficient algorithms with provable
optimality guarantees. For a submodular function f(S), two relevant problems that can be
formulated within this framework are (a) selecting a set of up to k elements in order to maximize
the function f(S), and (b) selecting the minimum-size set of inputs in order to achieve a given
bound α on f(S). The two problems are formulated as
maximizeS⊆V f(S)
s.t. |S| ≤ k
minimize |S|
s.t. f(S) ≥ α
(a) (b)
(2)
Submodularity implies that simple greedy algorithms are sufficient to approximate both
problems up to provable optimality bounds. For Problem 2(a), the algorithm is stated as follows:
9
U W
w1
w2
w3
w4
w5
U W
w1
w2
w3
w4
w5
u1
u2
u3
u4
u5
u6
u1
u2
u3
u4
u5
u6
X={1,2,5}independent X’={3,5}notindependent
Figure 3. A set X is independent in a transversal matroid if there is a matching where each
node in X is matched to exactly one node in W . In this example, the set X = {1, 2, 5} is
independent, since there is a matching in which each node in {wi : i ∈ X} is matched to exactly
one neighbor in U . Note that this matching is not unique: for example, w2 could be matched
to u2, u3, or u4. The set X ′ = {3, 5} is not independent, since nodes w3 and w5 can only be
matched to the same node u6.
1) Initialize the set S to be empty. Set i = 0.
2) If i = k, return S. Else go to 3.
3) Select the element v that maximizes f(S ∪ {v}). Update S as S ← S ∪ {v}. Increment i
by 1 and go to 2.
The greedy algorithm returns a set S ′ satisfying
f(S ′) ≥(
1− 1
e
)f(S∗),
where S∗ is the solution to (2(a)) [26].
The algorithm for approximately solving Problem 2(b) is similar:
10
1) Initialize the set S to be empty.
2) If f(S) ≥ α, return S. Else go to 3.
3) Select the element v that maximizes f(S ∪ {v}). Update S as S ← S ∪ {v}. Go to 2.
Letting S ′ denote the set returned by the second algorithm and S∗ denote the solution to (2(b)),
the two sets satisfy
|S ′||S∗|≤ 1 + log
{f(V )− f(∅)f(S ′)− f(S ′′)
},
where S ′′ denotes the value of the set at the iteration prior to termination of the algorithm [31].
Both algorithms terminate in polynomial time and only require O(n2) evaluations of the
objective function. The greedy algorithm can also be applied to the problem of minimizing a
supermodular function to obtain similar guarantees.
11
Sidebar 1: Applications of Submodularity
The intuitive “diminishing returns” nature of submodularity leads to inherent submodular
structures in a variety of application domains. One such problem consists of selecting a subset
of observations to maximize their information content, or equivalently reduce the uncertainty of
a random process [32][33]. The submodular structure of this problem arises because many of
the classical information-theoretic metrics used to evaluate uncertainty and information gathered,
such as entropy and mutual information, are inherently submodular. Practical applications include
sensor placement for monitoring temperature, structural health of buildings, and water quality
[32][34]. Detection of objections in a video frame has also been investigated using submodular
optimization methods [35].
Influentialuser
Influentialuser
Figure 4. Influence maximization problem in social network. The goal of the problem is to select
a subset of users, as part of a marketing campaign for example, who are likely to influence other
users, creating word of mouth that propagates through the network. The problem of selecting
a set of users that maximize the level of influence is known to be submodular for a variety of
relevant influence models [10].
Submodular optimization is an important tool in social network analysis, due to the need
for scalable algorithms over network datasets with millions of nodes. A seminal result established
12
that identifying the most influential nodes in a social network can be formulated as a monotone
submodular maximization problem [10] under several widely accepted influence models (Figure
4). Related problems, such as selecting sets of influential blogs, have also been studied under the
submodular framework [36]. An additional problem in social networking consists of identifying
the most likely set of links in a social network, based on an observed diffusion process. This
learning and estimation problem has also been shown to be submodular [37].
Document summarization is the problem of selecting a subset of keywords to best
describe a text. Intuitively, the descriptive power of a set of keywords might be expected to
be submodular; as more keywords are added to the set, the amount of information gained from
each additional keyword diminishes. This intuition was established rigorously by the discovery
that some standard information content metrics are submodular as a function of the set of
keywords [38]. This led to the development of submodular algorithms for keyword selection,
which provided better summarization performance than the current state of the art.
Finally, submodularity frequently arises when considering the economic utility of rational
entities. Submodularity was proposed in a game-theoretic context under the class of convex
cooperative games [39]. In a convex game, a coalition of cooperating players which is stable
(no player has an incentive to leave the coalition) can be computed in polynomial time.
13
Sidebar 2: Graph Matching
Consider a scenario where n users, denoted {u1, . . . , un}, are choosing from a set of m
items denoted {w1, . . . , wm}. Each user ui has a set of “desirable” items Wi. Each user can
receive at most one item, and each item can be given to at most one user. The goal in this
scenario is to choose a set of (user, item) pairs to maximize the number of users who receive
items that are desirable to them.
This problem can be modeled as a bipartite graph. A bipartite graph is a graph whose
vertex set V can be partitioned into disjoint subsets U and W such that all edges in the graph
are between U and W (Figure 5). In this scenario, the set U corresponds to users, while the set
W corresponds to items, and an edge (ui, wj) exists if item wj is desirable to user ui.
U Wu1
u2
u3
u4
u5
u6
w1
w2
w3
w4
Figure 5. A bipartite graph. Highlighted edges form a maximal matching, in which u1 is
matched to w1, u2 is matched to w2, u4 is matched to w3, and u6 is matched to w4. Note that
the matching is not unique, for example, u5 could be matched to w3 instead.
A matching on a bipartite graph is a subgraph in which each node has degree at most
14
one, so that each node in U is “matched” to at most one node in W . Equivalently, a matching
can be viewed as a one-to-one map from U into W . A maximal matching is a matching in
which no additional edges can be added from U into W while satisfying the requirement that
the degree of each node is bounded by one. All maximal matchings can be shown to have the
same cardinality [40]. A matching in which all nodes are matched is a perfect matching.
The problem of choosing the mapping between users and items is equivalent to finding
a maximal matching on the graph. This problem can be solved in polynomial-time using known
methods such as the Edmonds augmenting path algorithm [41]. Other polynomial-time solvable
problems include maximum-weight matching (maximizing the utility of the set of users when
each user ui attaches benefit bij to item wj) and maximum-weight maximum matching (selecting
a maximal matching that maximizes the total benefit) [40].
Dynamics of Networked Systems
A networked system can be modeled as a graph, in which each node of the graph
represents one component of the system. A node could represent a single bus in a power system,
a gene or protein in a regulatory network, or one vehicle in a transportation network. Each
node can be assigned an index i ∈ {1, . . . , n} in a system of n nodes total. Each node has a
time-varying internal state xi(t) ∈ R.
The interactions between networked system components are represented by edges in
the graph, with two nodes sharing an edge if the dynamics of the corresponding components
are coupled to each other. Examples include nodes in communication or formation control
networks that are within radio range of each other, and hence can directly share information,
15
buses connected by transmission lines, and genes that directly regulate each other. Together,
the nodes and edges form a graph G = (V,E), where V = {1, . . . , n} and E = {(i, j) :
i influences the dynamics of j}.
The set of nodes that have an edge incoming to node i is denoted Nin(i) = {j : (j, i) ∈
E}, and is interpreted as the set of nodes that are influenced by i. Similarly, the set of nodes
that are influenced by node i is denoted Nout(i) = {j : (i, j) ∈ E}. The in-degree (respectively,
out-degree) is defined by |Nin(i)| (respectively, |Nout(i)|). When the graph is undirected, with
(i, j) ∈ E implying (j, i) ∈ E, the in-degree and out-degree are equal and are referred to as the
degree di, while N(i) , Nin(i) = Nout(i) is the neighbor set of node i.
If there is a set of edges (i0, i1), (i1, i2), . . . , (im−1, im) with i = i0 and j = im, then this
set of edges forms a path from i to j. A graph is strongly connected if there is a path between
any pair of nodes.
In some settings, the interactions between neighboring nodes can be treated as a linear
coupling, in which the dynamics of node i are given by
xi(t) = Wiixi(t) +∑
j∈Nin(i)
Wijxj(t) (3)
where Wij are nonzero weights, when the graph is directed, and
xi(t) = Wiixi(t) +∑j∈N(i)
Wijxj(t) (4)
when the graph is undirected.
A variety of methods have been developed to control or influence networked systems.
These methods include creating or removing links in order to shape the network dynamics
[42], providing incentives in social networks [43], and developing distributed strategies for each
16
individual agent to reach a shared goal [44]. This article considers control techniques in which
a subset S of nodes, denoted input nodes, have their state values determined directly by an
external entity, such as a remote operator, drug intervention, or an external location signal. The
states of the input nodes are then treated as external control signals, resulting in a model
xf (t) = Afxf (t) +Bfu(t)
where xf (t) and u(t) are the state vectors of the non-input and input nodes, respectively.
The choice of input nodes is known to impact the performance and controllability of
networked systems. The impact of the set of input nodes on the robustness of the networked
system to noise was studied in [19], [20]. The selected input nodes also determine the rate at
which the network dynamics converge to the desired steady-state value [21], [45], [23]. Finally,
the controllability of the networked system, defined as the ability of the input nodes to drive the
network from any initial state to any final state, depends on which nodes are chosen as inputs
[15], [11]. Three input selection problems that can be solved within the submodular optimization
framework are:
Problem 1: How to select a set of up to k input nodes in order to maximize a performance or
controllability metric (analogous to Eq. (2(a)))?
Problem 2: How to select the minimum-size set of input nodes in order to ensure that the system
satisfies a given bound on performance and controllability (analogous to Eq. (2(b)))?
Problem 3: How to select a set of up to k input nodes in order to maximize a performance metric
while guaranteeing controllability?
Heuristics such as selecting high-degree nodes to act as inputs may lead to suboptimal
17
solutions to each of the three problems [46], [11]. The importance of the input nodes motivates
the development of an analytical framework for input selection, which will be presented in the
following sections.
18
Sidebar 3: Random Walks on Graphs
A random walk models the behavior of a particle moving at random over a graph G =
(V,E) according to a stationary probability distribution. Formally, a random walk is a discrete-
time random process X[k], with X[k] ∈ V . A random walk is defined by its transition matrix
P , where Pij = P (X[k] = j|X[k − 1] = i) represents the probability that the walk transitions
(takes a step) from location i to location j (Figure 6). A transition matrix is stochastic, meaning
it has nonnegative entries and rows that sum to 1.
The behavior of the random walk can be quantified by metrics including the stationary
distribution and mixing time. The stationary distribution is the steady-state probability distribution
of the walk, and is equal to the solution π of πP = π, where P is the transition matrix [47]. If
the graph is connected and the greatest common divisor of the cycle lengths is 1 (equivalently,
the walk is irreducible and aperiodic), then the stationary distribution π is unique and the walk
converges in probability to the stationary distribution. The mixing time is the rate at which
the random walk converges to the stationary distribution, and can be quantified through the
eigenvalues of the transition matrix [48].
The behavior of a random walk on a graph can be analyzed through statistics including
the hitting, commute, and cover times (Figure 7). The hitting time H(v, S) is the expected time
for a random walk to reach any node in a set of vertices S from a given initial location v. The
commute time κ(v, S) is the expected time for a random walk originating at v to reach any node
in a set S and then return to v. The cover time C(v) is the expected time for a random walk
originating at v to reach all vertices in the graph. These statistics are known to have connections
19
n1
n2n3 n4
n5 n6n7
P52P53
P55
Figure 6. Illustration of a random walk with transition matrix P on a graph. Pij denotes the
probability of a transition from node ni to node nj .
v S
Hittingtime:3Commutetime:8Covertime:9
Figure 7. Hitting, commute, and cover times of a random walk on a graph. The times shown
represent a single sample path of the walk, indicated by green arrows. The overall hitting,
commute, and cover times are obtained by averaging over all sample paths.
to physical quantities including the effective resistance of the graph [49], the rate of diffusion on
the graph [50], and the performance of some distributed communication protocols such as gossip
and query processing [51], [52]. This article describes the connection between random walks
and the performance (robustness to noise and convergence rate) of networked control systems.
20
Submodularity for Performance: Robustness to Noise
Networked systems operate in inherently lossy and insecure environments, which lead
to multiple sources of errors. Noise in communication links can cause incorrect computation
of state updates, and hence error in the state dynamics of the nodes. Physical disturbances are
common, as are modeling errors due to simplified models of complex, nonlinear node dynamics
and interactions. Errors can also be caused by malicious adversaries through intelligent attacks
(such as denial-of service or false data injection). This section considers input selection to
minimize the impact of noise for two classes of system dynamics, namely consensus dynamics
and Kalman filtering. Consensus dynamics are widely used in formation control [1], distributed
estimation [53], and synchronization applications [54], while distributed Kalman filtering is a
standard approach for estimation and filtering over networks [55].
Input Selection for Robustness to Noise
Consider a group of unmanned vehicles who exchange information over a communication
network. The goal of the vehicles is to follow a trajectory maintained by an external signal (e.g.,
from a remote operator) sent to the input nodes while maintaining a formation, in a noisy and
bandwidth-limited wireless network.
Under the consensus-based approach in the presence of noise, the input nodes maintain
constant states (such as the desired heading or velocity) while the non-input nodes have dynamics
xi(t) = −∑j∈N(i)
(xi(t)− xj(t)) + wi(t), (5)
where wi(t) is a zero-mean white noise process [56], [57], [58]. The network topology is assumed
21
to be undirected in this section. An advantage of this approach is that, in the absence of any
noise, the states of the nodes will converge to the input node states. Furthermore, the dynamics
only require each node to communicate with its one-hop neighbors, reducing the communication
overhead required and making the system robust to topology changes. Other applications of the
dynamics (5) include location estimation, time synchronization, and opinion dynamics in social
networks. In order to further analyze the consensus dynamics with set of input nodes S, consider
the graph Laplacian matrix defined by
Lij =
−1, (i, j) ∈ E, i /∈ S
di, i = j, i /∈ S
0, else
(6)
The Laplacian matrix can be decomposed as
L =
Lff Lfl
0 0
,
where Lff represents the impact of the non-input nodes on each others’ state values, Lfl is the
impact of the input node states on the non-input nodes, and the remaining zero entries reflect
the fact that the input nodes maintain constant state values. The system dynamics can be written
in vector form as
xf (t) = −Lffxf (t)− Lflxl(t) + w(t).
One metric for the robustness of the system to noise is the H2 norm, which is equal to
the asymptotic mean-square deviation from the desired state. Letting x∗ = limt→∞ xf (t), x∗ is
a random variable with covariance matrix Σ given by the solution to the Lyapunov equation
LffΣ + ΣLff = I.
22
Solving this equation for Σ yields Σ = 12L−1ff [19]. The metric R(S) , trace(L−1ff ) therefore
quantifies the mean-square error due to noise in the node states.
Random Walks and Error in Networked Systems
As a first step towards developing a submodular optimization approach to minimizing
errors due to noise, a connection will be established between error due to noise and the statistics
of a random walk on the network, specifically, the commute time. For details on random walks,
see “Random Walks on Graphs”.
vu*=1 v2* v3*
v4* v5*
v6*=0 v7*v8*=0
𝑣"# = 1 & 𝑃#( + 𝑣"# & 𝑃#* + 𝑣"+ & 𝑃#+𝑣", = 1
𝑣"*
𝑣"+𝑣"-
𝑣".𝑣"/ = 0
𝑣"1 = 0
(a) (b)
Figure 8. Description of the connection between the inverse Laplacian and a random walk on
the graph. (a) Illustration of the vector v∗ used to compute the error due to noise at node u,
equal to (L−1ff )uu, when S = {6, 8}. (b) Connection to a random walk. The variable vi is equal
to the probability that a walk originating at i reaches u before S under a random walk with
transition matrix P . P21, P23, and P25 denote the transition probabilities of the walk. It can be
shown using the Maximum Principle that v∗ = v.
23
Observe that the u-th diagonal element of L−1ff , denoted (L−1ff )uu, can be obtained as
follows. Define vectors v∗ and J as solutions to the equation Lv∗ = J satisfying v∗u = 1, Ji = 0
for all i ∈ V \ (S ∪ {u}), and v∗i = 0 for all i ∈ S (Figure 8(a)). Solving this equation for the
remaining entires of v∗ and J yields (L−1ff )uu = 1/Ju. Equivalently,
(L−1ff )uu =
∑j∈N(u)
(1− v∗j )
−1 .The error variance of node u, equal to (L−1ff )uu, can therefore be expressed as a function of v∗.
Note that
v∗i =∑j∈N(i)
|Lij|Lii
v∗j (7)
for all i /∈ (S ∪ {u}). On the other hand, consider a random walk with transition matrix Pij =
|Lij|/Lii|, and define vi(S, u) to be the probability that a random walk originating at i reaches
u before S (Figure 8(b)). By construction, vi(S, u) also satisfies (7). In fact, it can be shown
that vi(S, u) = v∗i for all i using the maximum principle from harmonic analysis [49], leading
to the following result.
Theorem 1 ([46]): The variance of the error due to noise at node u, (L−1ff )uu, is
proportional to the commute time κ(S, u).
The connection between the error due to noise and commute time can also be developed
using the graph effective resistance. This approach is not presented here due to space constraints,
but can be found in [59]. Further reading on the connection between error due to noise and
effective resistance can be found in [20].
24
Supermodularity of Error Due to Noise
This subsection establishes supermodularity of the robustness to noise by exploiting the
connection to the commute time. The key result is the following.
Theorem 2 ([46]): The commute time κ(S, u) is a supermodular function of the set S.
As a corollary, the mean-square error metric R(S) is supermodular as well. A sketch of
the proof of Theorem 2 is as follows. First, note that the goal is to show that for any S ⊆ T
and any v /∈ T ,
κ(S, u)− κ(S ∪ {v}, u) ≥ κ(T, u)− κ(T ∪ {v}, u).
Let UjSu be a random variable, equal to the time for a walk starting at a node j to reach S and
then reach u. Let τj(S) denote the event that a walk reaches j before S.
The differences in the commute times can then be rewritten using the expression
and K is an upper bound determined by the norm of x(0) [45]. The function gij(S) is the
probability that a walk starting at node i reaches node j after τ steps, while hi(S) is the
probability that a walk starting at node i does not reach the input set at step τ . The choice of
input set impacts the values of gij(S) and hi(S) because each input node is an absorbing state
of the walk.
In order to develop a submodular approach to selecting input nodes for smooth
convergence, it suffices to prove supermodularity of gij(S) and hi(S). The supermodular structure
31
arises because the set of input nodes S represents a set of absorbing states of the walk. If a
walk reaches a node in S at step τ ′ < τ , then the walk remains at that node and does not reach
j ∈ V \ S at step τ . Thus the increment gij(S)− gij(S ∪ {v}) is equal to the probability that a
walk starting at i reaches v and j by time τ , but does not reach the set S. The inequality
gij(S)− gij(S ∪ {v}) ≥ gij(T )− gij(T ∪ {v}) (11)
can be shown by considering separate cases of S and T , illustrated in Figure 12.
j
i
T
Sv i
j
T
Sv i
j
T
Sv
(a) (b) (c)
Figure 12. Three cases to illustrate supermodularity of the convergence error as a function of
the input set. (a) The walk reaches the set S, and hence both sides of (11) are zero. (b) The
walk reaches v but not S, implying that both sides of (11) are equal. (c) The walk reaches v
and T \ S but not S, implying that (11) holds with strict inequality.
If the walk reaches S (Figure 12(a)), then both sides of (11) are automatically zero, since
the probability that the walk continues to node j after reaching S is zero. Similarly, if the walk
does not reach T , then the input set does not impact the walk and hence both sides of (11) are
equal (Figure12(b)). In the case shown in Figure 12(c), the walk reaches T and v but not S
within τ steps. Hence the left-hand side of (11) is positive, while the right-hand side is zero,
implying that (11) holds with strict inequality.
Combining these arguments and the composition rules for supermodular functions yields
32
the following main result.
Theorem 4 ([45]): The convergence error bound
ft(S) =∑i∈V \S
∑j∈V \S
gij(S)p + hi(S)p
is supermodular as a function of the input set S.
The supermodularity of the convergence error implies that efficient algorithms can be
developed for selecting input nodes to minimize the deviation of the intermediate node states.
Moreover, Theorem 4 implies that generalizations of the convergence error, such as the integral∫∞0ft(S) dt, are supermodular functions of the input set [45].
Input Selection in Dynamic Networks
The discussion so far has implicitly assumed that the network topology G = (V,E) is
fixed over time. The topologies of networked systems, however, evolve over time. The model
for the network topology dynamics naturally depends on the cause of these variations. In many
cases, however, the submodular framework extends naturally to dynamic topologies.
One source of topology changes is random failures of nodes or links in an otherwise
static topology. Node failures may occur due to hardware failures, or participants dropping out
of a social network, while link failures often occur due to communication over lossy wireless
channels. In both cases, for a given performance metric f(S), the effect of the topology can be
quantified as the expected value of the metric, f(S) = Eπ(f(S)), where π denotes the probability
distribution on the network topology due to node and link failures. The following result enables
extending the submodular design approach to network topologies with random failures.
33
Lemma 1 ([10]): If f(S) is submodular (respectively, supermodular), then the function
f(S) = Eπ(f(S)) is a submodular (respectively, supermodular) function of S.
The proof can be seen from the fact that f(S) is a nonnegative weighted sum of submodular (or
supermodular) functions. While submodularity of f(S) enables provable guarantees for simple
greedy algorithms, the complexity of evaluating the function f(S) is worst-case exponential,
since all possible network topologies may have nonzero probability. Monte Carlo methods or
approximations to specific cost functions may be used in this case [46].
The network topology may also undergo changes caused by switching between predefined
topologies. Switching topologies are common in formation maneuvers, where different topologies
are used to change the coverage area of the formation and avoid obstacles [66], [67].
The effect of the switching can be modeled as a set of topologies {G1, . . . , GM}. Two
relevant metrics are the average and worst-case performance. The average case performance,
given as favg(S) = 1M
∑Mi=1 f(S|Gi), inherits the submodular structure of the objective function
f(S), similar to the case of random failures.
The worst-case performance is formulated as fworst(S) = max {f(S|Gi) : i = 1, . . . ,M},
and unlike favg is not supermodular as a function of S. The problem of selecting a minimum-
size input set to achieve a desired bound on the worst-case performance can, however, be
approximated with provable optimality guarantees by using the equivalent formulation [68],
[46], [45]
minimize |S|
s.t. maxi=1,...,M f(S|Gi) ≤ α
⇔minimize |S|
s.t. 1M
∑Mi=1 max {f(S|Gi), α} ≤ α
(12)
The function max {f(S), c} is supermodular whenever f(S) is a decreasing supermodular
34
function and c is a real constant, and hence the equivalent problem formulation defines a
supermodular optimization problem.
Submodularity and Controllability
A networked system is controllable if it is possible to drive the node states x(t) from any
initial values x(0) to any desired final values x(T ) in a finite time T . The problem of selecting
input nodes to guarantee controllability has received significant attention, including results on
controllability of consensus networks [15], [21], [69], networks with known parameters [70], and
topologies with known and unknown parameters (structured systems) [11], [18], [71], [72]. This
section presents submodular optimization techniques for selecting input nodes for controllability.
Controllability of Networked Systems
Conditions for controllability of linear systems have been studied since the 1960s, when
Kalman’s controllability criteria were presented. It was shown that a linear system
x(t) = Ax(t) +Bu(t) (13)
with x(t) ∈ Rn is controllable if and only if the controllability matrix C =
(B AB A2B · · · An−1B) has full rank.
In a networked system, the input nodes affect the matrix C by determining the columns
of the B matrix. If the dynamics of the system in the absence of any inputs are given by
x(t) = Wx(t), then selecting an input set S results in A = (Wij : i, j ∈ V \ S) and B = (Wij :
i ∈ V \ S, j ∈ S). Based on this insight, the problem of selecting a minimum-size set of input
35
nodes to ensure controllability can be formulated as
minimize |S|
s.t. r(S) = n
(14)
where r(S) = rank(C(S)) and C(S) is the controllability matrix when the input set is S. The
following result leads to submodular approaches to selecting input nodes for controllability.
Lemma 2 ([70]): The function r(S) is a monotone submodular function of S.
By Lemma 2, a simple greedy algorithm suffices to select a set of input nodes with
provable bounds on the cardinality of the set. Indeed, by applying the bounds on the greedy
algorithm for submodular cover, it follows that the set S selected by the greedy algorithm satisfies
|S||S∗|≤ 1 + log n,
where S∗ denotes the optimal solution. Furthermore, it can be shown that this is the best bound
that can be achieved unless the P=NP conjecture from complexity theory holds [71].
In addition to the controllability of the system, the controllability matrices provide insight
into the amount of energy that must be exerted to control a networked system. One such
controllability matrix is the controllability Gramian, defined as the positive semidefinite solution
WS to
AWS +WSAT +BSB
TS = 0.
The H2 norm of the system is a weighted trace of the controllability Gramian, so that ||H||22 =
tr(XWSXT ) for some matrix X . The trace, in turn, is a modular function of the set of input
nodes [70].
36
An additional energy-related metric is the trace of the inverse of the controllability
Gramian. This metric is proportional to the energy needed on average to steer the system from
the initial operating point to the final, desired state. The trace of the inverse of the controllability
Gramian is a monotone decreasing and supermodular function of the input set S [70].
Structural Controllability
In the preceding analysis, it was assumed that all parameters, such as interaction weights,
between nodes are known a priori. In many systems of interest, such as biological networks, the
parameters cannot be observed directly, or are estimated with errors or uncertainties. In such
systems, controllability can still be analyzed by considering the structural rank of the system.
Definition 1 ([73]): The structural rank of a system is defined as the maximum rank of
the controllability matrix over all values of the weights Wij in (3). A system satisfies structural
controllability if the structural rank of the controllability matrix is equal to the number of non-
input nodes.
Choosing the structural rank as the maximum achievable rank may seem optimistic. It
can, however, be shown that any set of parameters Wij achieve the structural rank, except when
the weights are chosen from a set that has Lebesgue measure zero [73]. Stricter conditions have
also been formulated; in [74], conditions for strong structural controllability, which implies
controllability for any nonzero values of the free parameters, are presented. Controllability
conditions for linear descriptor systems, which have a combination of free and fixed parameters,
are discussed in [75]. Furthermore, other structural conditions have been proposed, including
disturbance rejection properties [76], which can be relaxed to matroid constraints [77]. In what
37
follows, however, the analysis focuses on structural controllability as in Definition 1.
An advantage of structural controllability is that it can be characterized using properties
of the network graph, enabling a common set of techniques to be used to analyze systems in
different application domains. To motivate one necessary condition for structural controllability,
consider the graph shown in Figure 13.
n1
n2
n3 n6
n4
n5
Inputnode
Figure 13. The accessibility condition. The nodes n4 and n5 are not accessible from the input
node n6, and hence the graph is not controllable.
By inspection, the A and B matrices from Eq. (13) arising from this graph (which has
input node n6) are of the form
A =
0 ∗ ∗ 0 0
∗ 0 0 0 0
0 0 0 0 0
0 0 0 0 ∗
0 0 0 ∗ 0
, B =
0
∗
∗
0
0
and hence the states of nodes n4 and n5 are not controllable. From the graph-theoretic viewpoint,
the nodes n4 and n5 are not connected to the input node, and hence cannot be controlled from
38
that node. This motivates the accessibility property, defined as follows.
Definition 2: A node in a networked system is accessible if there is a path from an input
node to that node. The system satisfies accessibility if all nodes are accessible.
Accessibility is a necessary condition for controllability. For the next controllability
condition, consider the example of Figure 14.
n1
n2
n3 n6
n4
n5
Inputnode
Figure 14. Network that contains a dilation. The system is not controllable because the two
nodes {n1, n3} have only one neighbor, namely the input node n6. In this case, N(A) = {n6}
and |A| > |N(A)|.
In the graph of Figure 14, the nodes n1 and n3 both have exactly one neighbor, the input
node n6. Hence the state dynamics of node n1 and n3 satisfy x1(t) = αx3(t) for some constant
α ∈ R, and the states x1(t) and x3(t) always lie in an affine subspace of R2 that is determined
by the initial state and α. It is therefore impossible to drive x1(t) and x3(t) to any arbitrary
values, implying that controllability does not hold. We generalize this property to the notion of
dilation-freeness.
39
Definition 3: A network is dilation-free if, for any set of nodes A ⊆ V , |N(A)| ≥ |A|,
where N(A) = ∪i∈ANin(i) (in words, the number of neighbors of A is at least as large as the
set A).
Dilation-freeness can be interpreted by the following intuition. For a set of m nodes, in
order for those nodes to be driven to any arbitrary state, at least m degrees of freedom would
be needed. Otherwise, the inputs received by any two nodes would satisfy a linear relationship,
implying that the vector of node states x(t) would lie in a subspace of Rn. In Figure 14, the set
A = {n1, n3} and N(A) = {n6}, hence violating dilation-freeness.
Both accessibility and dilation-freeness are necessary conditions for structural controlla-
bility. It can also be shown that the converse is true.
Theorem 5 ([73]): If a networked system satisfies accessibility and dilation-freeness, then
the system is structurally controllable.
The dilation-free property also has a connection to matching theory (see Graph Matchings
sidebar), which can be understood using the Hall Marriage Theorem.
Theorem 6 (Hall Marriage Theorem [40]): For any bipartite graph, there exists a perfect
matching if and only if each set A ⊆ V satisfies |N(A)| ≥ |A|.
From Theorem 6, the dilation-free property is equivalent to the existence of a perfect
matching in the bipartite graph G = (U,Z,E), where Z = V \ S (the set of non-input nodes),
U is the set of all nodes in the network, and the edge (ui, zj) exists if (i, j) ∈ E.
40
n1
n2 n3
n4
n6 n5
n1
n2
n3
n4
n5
n1
n3
n4
n6
N(V\S)
n6
V\Sn1
n2
n3
n4
n5
n1
n3
n4
n6
N(V\S)
n6
V\S
Figure 15. Mapping controllability to a matching constraint. The first step is to map the
network graph to a bipartite representation, and then construct a maximal matching. Since there
is a matching from N(V \ S) into (V \ S) in which all nodes in V \ S are matched, the graph
is controllable from input set S = {n2, n5}.
In the case where the graph is strongly connected, the minimum-size set of input nodes to
guarantee structural controllability can be chosen based on a graph-matching algorithm. Under
the algorithm, a maximum matching on the graph is computed using a technique such as the
Hungarian algorithm [40]. All nodes that are left unmatched under the maximum matching are
then chosen as inputs. The connection between graph matchings and controllability is illustrated
in the example of Figure 15.
The matching condition on controllability can be expressed as a matroid constraint
through the following analysis. The mapping to matroids is a step towards developing joint
input selection algorithms for performance and controllability.
Consider the problem of selecting a feasible set of non-input nodes. If there is a matching
in which all of these non-input nodes are matched, then the graph is controllable. Define a set
41
I by
A ∈ I ⇔ There exists a matching where A is matched.
The following result maps controllability to a matroid constraint.
Theorem 7 ([18]): The tuple (V, I) defines a transversal matroid.
Controllability can therefore be expressed as a matroid constraint V \ S ∈ I. Many of
the properties of this matroid have physical interpretations. The bases of the matroid correspond
to the minimum-size input sets. For any set T ⊆ V , the rank function r(T ) is equal to the
maximum number of nodes in T that are matched, and hence is equivalent to the number of
nodes that are controllable when the input set is S = V \ T .
A related problem to selecting a minimum-size set of input nodes for controllability is
determining how effective a given set of input nodes is at controlling a graph. One graph-based
controllability metric, denoted as the graph controllability index (GCI), is defined by [18]
GCI(S) = max {|V ′| : Graph (V ′, E(V ′)) is controllable from S}
where E(V ′) is the set of edges in E that are between nodes in V ′. As an example, the GCI of the
graph shown in Figure 15 is 6, since all non-input nodes are matched under a maximal matching.
If S = {n5}, then the maximum-cardinality matching that can be obtained is 4, implying that a
total of five nodes (one input and four non-input nodes) are controllable.
In order to compute the GCI, observe that the set of nodes that are controllable can be
decomposed into the set of input nodes and the set of controllable non-input nodes. The set of
input nodes has cardinality |S| by definition. The set of controllable non-input nodes, by the
preceding discussion, has cardinality r(V \ S). Hence the graph controllability index can be
42
written as
GCI(S) = r(V \ S) + |S|,
which is the sum of a matroid rank function and the cardinality function and hence is monotone
increasing and submodular.
Minimizing Controller Energy
Controllability refers to the ability of the controller to steer the network states to any
desired values in a finite time by providing arbitrary input signals. In practice, however, an
arbitrary input signal may require high levels of energy, making control from a given input set
infeasible even if the controllability condition is satisfied. The minimum control effort problem
for a given input set is formulated as
minimize∫ t1t0u(t)Tu(t) dt
u(t) : t ∈ [t0, t1]
s.t. x(t) = Ax(t) +Bu(t), t ∈ (t0, t1]
x(t0) = x0, x(t1) = x1
(15)
The solution to this optimization problem is characterized by the controllability Gramian WS ,
and results in a minimum energy given by
(x1 − eA(t1−t0))TΓ(t0, t1)−1(x1 − eA(t1−t0)x0)
where Γ(t0, t1) is the controllability matrix. The impact of the choice of input nodes on the
controllability Gramian can be seen for the case where the B matrix is diagonal, so that each
incoming signal impacts exactly one input node. In this case, the matrix Γ can be written as∑i∈S Γi, where Γi =
∫ t1t0eAtδie
AT t dt and δi is a matrix with a 1 in the (i, i)-th entry and zeros
43
elsewhere. The problem of selecting a minimum-size set of inputs to ensure that the controller
energy is below a desired value R can then be formulated as
minimize |S|
s.t. vTΓ−1v ≤ R
(16)
where v = (x1− eA(t1−t0)x0). A submodular function that is arbitrarily close to the constraint in
(16) is given by
fε(S) , vT (Γ + εI)−1v + ε
n−1∑i=1
vTi (Γ + ε2I)−1vi,
where v1, . . . , vn−1 are an orthonormal basis for the null space of v.
Theorem 8 ([24]): For any ε > 0, the function fε(S) is supermodular as a function of
S.
The results of this section imply that, for a variety of systems, selecting a set of input
nodes to satisfy controllability can be formulated as a submodular optimization problem, implying
the existence of computationally efficient and provably optimal input selection algorithms for
controllability. Application domains include selecting a subset of leaders in a leader-follower
formation network in order to ensure that any specified trajectory can be followed; choosing a
subset of genes to ensure that a cell can be steered to a desired final state; and selecting a set
of generators to control in order to stabilize a power system.
Putting It Together: Performance and Controllability
Consider the two networks in Figure 16. The network on the left can be shown to satisfy
controllability (see the section“Submodularity and Controllability”), however, the chosen input
node is distant from the remaining network nodes, and hence suboptimal for performance metrics
44
Inputnode
Inputnode
(a) (b)
n1 n2
n3
n4
n1 n2
n3
n4
Figure 16. Comparison of two possible input nodes. (a) The chosen input node satisfies
controllability but provides poor performance due to its distance from the non-input nodes.
(b) The input node n2 is centrally located but does not satisfy controllability, since there is a
dilation A ⊆ V \ {n2} with A = {n1, n3} and N(A) = {n2}.
such as smooth convergence, which rely on inputs reaching the remaining non-input nodes in
a timely fashion. On the other hand, in the network on the right, the input node is centrally
located but does not guarantee controllability. The potential conflicts between different design
requirements motivates the development of an analytical framework for joint input selection
based on performance and controllability.
The problem of selecting a set of up to k input nodes to ensure structural controllability
while maximizing a performance metric is given by
maximize f(S)
s.t. Network controllable from S
|S| ≤ k
(17)
The following theorem leads to a submodular approach to joint performance and
controllability.
45
Theorem 9: If the function f(S) is monotone, then there exists a matroid M such that
Problem (17) is equivalent to max {f(S) : S ∈M}.
If the function f(S) is submodular, then this problem is submodular maximization subject
to a matroid constraint. A modified version of the greedy algorithm suffices to approximate this
problem with provable optimality guarantees:
1) Initialize S = ∅.
2) If |S| = k, return S. Else go to 3.
3) Select v satisfying (S ∪ {v}) ∈M and v maximizes f(S ∪ {v}).
4) Set S ← (S ∪ {v}). Go to 2.
The greedy algorithm is guaranteed to achieve an optimality bound of 1/2 [78]. This optimality
bound can be improved to (1−1/e) through the continuous greedy algorithm [27]; the complexity
of the algorithm, however, precludes its use on large-scale networks.
An implicit assumption in (17) is that there exists a set of input nodes with cardinality k
that are sufficient for controllability. This condition can be relaxed through the graph controllabil-
ity index GCI(S), so that the optimization problem becomes max {f(S) + λGCI(S) : |S| ≤ k}
where λ is a nonnegative constant.
Finally, for systems with multiple performance and controllability constraints, the
46
constraints can be combined as
minimize |S|
s.t. f1(S) ≤ α1
...
fm(S) ≤ αm
⇔minimize |S|
s.t.∑m
i=1 min {fi(S), αi} ≥∑m
i=1 αi
(18)
and hence inputs can be selected jointly while maintaining provable optimality guarantees.
Numerical Studies
In order to illustrate the potential benefits of the submodular optimization approach to
control of networked systems, numerical studies of input selection for robustness to noise, smooth
convergence, and controllability are presented. For each problem, the following algorithms were
compared: (a) the submodular optimization approach, (b) selection of high-degree nodes to act
as inputs, (c) selection of average-degree nodes, and (d) random input selection. Each numerical
study was averaged over 50 independent trials.
Robustness to noise was evaluated in a network of n = 100 nodes. The network topology
was generated according to a geometric model, in which nodes are deployed uniformly at random
over a square region with width w = 1000m and an undirected edge exists between nodes i and
j if their positions are within 300m of each other. In this scenario, the submodular approach
requires 10-20 fewer inputs to achieve a desired error bound than the random and average degree
heuristics, and 20-40 fewer inputs than the maximum-degree input selection.
The numerical study of smooth convergence is shown in Figure 18, for a geometric graph
with n = 100 nodes, width 1400m, r = 250m, and p = 2. The convergence error arising from